An agent earns its keep only when it works without you

Before the team builds a tool out of momentum, the gap goes on paper: the PRD defines the problem, the users and the boundaries - and only on that basis do you decide whether to build at all.

Input: shape-notes.md from the discovery stage (or raw notes). Output: context/foundation/prd.md conforming to a locked schema - 10 sections for a greenfield project, 11 for a change in an existing system.

A 0-4 heuristic scores whether the notes are shaped: a checkpoint block, requirements in FR-NNN format, Given / When / Then scenarios, an explicit business rule. Below the threshold the skill warns outright: what is missing and what it costs the document. Generation happens in memory, with schema validation before anything lands on disk.

For a change in an existing system: a review of the current state, problem and motivation, personas, success criteria with "what must not break" guardrails, stories, functional and non-functional requirements, Non-Goals and Open Questions.

The skill is a document generator, not a discovery facilitator: it never invents domain decisions, business rules or success criteria. Every gap goes verbatim into Open Questions - a human resolves it, not the generator.

Review on every PR with no human in the loop: an agent in CI compares the diff against the plan the PR declares it implements, and leaves a durable audit trail - a report, inline comments, a verdict.

Input: a PR plus a plan - found by the context/changes/<id>/plan.md convention or a "Plan:" line in the PR description. Output: a reviews/impl-review.md report committed to the PR branch, inline comments on the changed lines and a summary comment with a table of dimensions.

Diff from the merge-base → three subagents in parallel: drift against the plan (MATCH / DRIFT / MISSING / EXTRA), safety and conformance with neighboring patterns, test coverage → verification commands from the plan → scores across seven dimensions → an APPROVED / NEEDS ATTENTION / REJECTED verdict.

Severity (CRITICAL / WARNING / OBSERVATION), impact, dimension, file:line location, detail and a proposed fix with explicit strength, cost and confidence. A limit of 10 findings - related problems get consolidated, and sorting by severity separates findings worth fixing from noise.

The skill asks nothing and edits no code: it reads, analyzes, reports. A missing plan in a PR means a neutral comment and an exit, not a failure. A REJECTED verdict blocks the check only in the workflow - with an override label after a human decision.

Register a name or schema that is load-bearing: changing it forces a coordinated update across its consumers. The registry makes future plan reviews flag contract-breaking changes automatically.

Input: a freshly created or newly recognized contract - a schema field, an API endpoint, a frontmatter format. Output: an H2 entry in docs/reference/contract-surfaces.md, which /10x-plan-review greps against the text of every future plan.

The surface name (the H2 heading is the literal grep target - fully qualified instead of a bare identifier), the canonical file:line definition, owners and consumers, and a breaking-change checklist ready to paste into a migration plan.

A four-question interview - the user writes the content, the skill suggests nothing upfront → an echo of the entry and confirmation → self-bootstrap of the file on first use → append at the end of the registry.

The file is append-only - revisions are made by a human, by hand, and that friction is deliberate. One entry per invocation. The entry bar: if a rename breaks something only in a local module, it is not a contract and no entry is created.

Execute an approved plan with nobody at the keyboard: the autonomous counterpart of /10x-implement for /goal sessions and headless runs. Trust is replaced by explicit policies, not by the agent's good will.

The goal condition is written down before launch: file scope, no weakening of tests, a turn limit. A no-interaction policy - nobody is watching, so every decision has a rule set in advance, and doubts are resolved by the conservative interpretation: fewer files, fewer behavior changes, closer to the letter of the plan.

Each phase's implementation is written by a delegated subagent, but the quality gates run in the main context: plan criteria → a deliberate-break check (the test must go red on deliberately broken code - otherwise it protects nothing) → the full test suite → commits only on green, one per phase.

A divergence between the plan and reality is classified: Minor (a moved file, a changed name) - adapted and reported in one line; Structural (a missing dependency, a different architecture) - a STOP block with resume instructions. When in doubt: Structural.

The goal evaluator reads only the transcript: every gate verdict, commit SHA and adaptation must be narrated in the response - a gate that passed silently does not exist. At most 2 repair attempts per gate, then STOP; manual steps return to the human as a closing checklist.